Arm NN Operators

Arm NN Operators

Arm NN supports operators that are listed in below table.

Arm NN supports a wide list of data-types. The main data-types that the Machine Learning functions support are the following:

• BFLOAT16: 16-bit non-standard brain floating point
• QASYMMU8: 8-bit unsigned asymmetric quantized
• QASYMMS8: 8-bit signed asymmetric quantized
• QUANTIZEDSYMM8PERAXIS: 8-bit signed symmetric quantized
• QSYMMS8: 8-bit signed symmetric quantized
• QSYMMS16: 16-bit signed symmetric quantized
• FLOAT32: 32-bit single precision floating point
• FLOAT16: 16-bit half precision floating point
• SIGNED32: 32-bit signed integer
• BOOLEAN: 8-bit unsigned char
• All: Agnostic to any specific data type

Arm NN supports the following data layouts (fast changing dimension from right to left):

• NHWC: Layout where channels are in the fastest changing dimension
• NCHW: Layout where width is in the fastest changing dimension
• All: Agnostic to any specific data layout

where N = batches, C = channels, H = height, W = width

Operator	Description	Equivalent Android NNAPI Operator	Backends	Data Layouts	Data Types
AbsLayer	Layer to perform absolute operation.	ANEURALNETWORKS_ABS	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	FLOAT16 FLOAT32 SIGNED32
			GpuAcc	All	FLOAT16 FLOAT32
ActivationLayer	Layer to simulate an activation layer with the specified activation function.	ANEURALNETWORKS_ABS ANEURALNETWORKS_ELU ANEURALNETWORKS_HARD_SWISH ANEURALNETWORKS_LOGISTIC ANEURALNETWORKS_PRELU ANEURALNETWORKS_RELU ANEURALNETWORKS_RELU1 ANEURALNETWORKS_RELU6 ANEURALNETWORKS_SQRT ANEURALNETWORKS_TANH GELU	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 FLOAT16 FLOAT32
AdditionLayer	Layer to add 2 tensors.	ANEURALNETWORKS_ADD	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
ArgMinMaxLayer	Layer to calculate the index of the minimum or maximum values in a tensor based on an axis.	ANEURALNETWORKS_ARGMAX ANEURALNETWORKS_ARGMIN	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32 SIGNED64
			CpuAcc	All	QASYMMU8 QASYMMS8 SIGNED32 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 SIGNED32 FLOAT16 FLOAT32
BatchMatMulLayer	Layer to perform batch matrix multiplication.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	FLOAT32 QASYMMS8
			GpuAcc	All	FLOAT32 QASYMMS8
BatchNormalizationLayer	Layer to perform batch normalization.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	FLOAT32 FLOAT16
			GpuAcc	NHWC NCHW	FLOAT32 FLOAT16
BatchToSpaceNdLayer	Layer to perform a batch to space transformation.	ANEURALNETWORKS_BATCH_TO_SPACE_ND	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
BroadcastToLayer	Layer to broadcast a tensor to a given size.	N/A	CpuRef	N/A	N/A
			CpuAcc	N/A	N/A
			GpuAcc	N/A	N/A
CastLayer	Layer to cast a tensor to a type.	ANEURALNETWORKS_CAST	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QSYMMS8 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32 SIGNED64
			CpuAcc	All	FLOAT32 FLOAT16 QASYMMS8 QASYMMU8 SIGNED32 SIGNED64
			GpuAcc	All	FLOAT32 FLOAT16 QASYMMS8 QASYMMU8 SIGNED32 SIGNED64
ChannelShuffleLayer	Layer to reorganize the channels of a tensor.	ANEURALNETWORKS_CHANNEL_SHUFFLE	CpuRef	All	FLOAT16 FLOAT32 QSYMMS8 QASYMMS8 QASYMMU8
			CpuAcc	All	QASYMMS8 QASYMMU8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMS8 QASYMMU8 FLOAT16 FLOAT32
ComparisonLayer	Layer to compare 2 tensors.	ANEURALNETWORKS_EQUAL ANEURALNETWORKS_GREATER ANEURALNETWORKS_GREATER_EQUAL ANEURALNETWORKS_LESS ANEURALNETWORKS_LESS_EQUAL ANEURALNETWORKS_NOT_EQUAL	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 BOOLEAN QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	All
			GpuAcc	All	All
ConcatLayer	Layer to concatenate tensors along a given axis.	ANEURALNETWORKS_CONCATENATION	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
ConstantLayer	Layer to provide a constant tensor.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16 SIGNED32
			CpuAcc	All	All
			GpuAcc	All	All
ConvertFp16ToFp32Layer	Layer to convert Float16 tensor to Float32 tensor.	N/A	CpuRef	All	FLOAT16 FLOAT32
			CpuAcc	All	FLOAT16 FLOAT32
			GpuAcc	All	FLOAT16 FLOAT32
ConvertFp32ToFp16Layer	Layer to convert Float32 tensor to Float16 tensor.	N/A	CpuRef	All	FLOAT16 FLOAT32
			CpuAcc	All	FLOAT16 FLOAT32
			GpuAcc	All	FLOAT16 FLOAT32
Convolution2dLayer	Layer to compute a convolution operation.	ANEURALNETWORKS_CONV_2D ANEURALNETWORKS_GROUPED_CONV_2D	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
			GpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
Convolution3dLayer	Layer to compute a 3D convolution operation.	N/A	CpuRef	NDHWC	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	N/A	N/A
			GpuAcc	N/A	N/A
DebugLayer	Layer to print out inter layer tensor information.	N/A	CpuRef	All	BFLOAT16 BOOLEAN FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16 SIGNED32
DepthToSpaceLayer	Layer to perform Depth to Space transformation.	ANEURALNETWORKS_DEPTH_TO_SPACE	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
DepthwiseConvolution2dLayer	Layer to compute a deconvolution or transpose convolution.	ANEURALNETWORKS_DEPTHWISE_CONV_2D	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	NHWC NCHW	FLOAT16 FLOAT32 SIGNED32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
			GpuAcc	NHWC NCHW	FLOAT16 FLOAT32 SIGNED32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
DequantizeLayer	Layer to dequantize the values in a tensor.	ANEURALNETWORKS_DEQUANTIZE	CpuRef	All	QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	All	FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS QSYMMS8 QSYMMS16
			GpuAcc	All	FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS QSYMMS8 QSYMMS16
DetectionPostProcessLayer	Layer to generate the detection output based on center size encoded boxes, class prediction and anchors by doing non maximum suppression (NMS).	ANEURALNETWORKS_DETECTION_POSTPROCESSING	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT32
DivisionLayer	Layer to divide 2 tensors.	ANEURALNETWORKS_DIV	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	FLOAT16 FLOAT32
			GpuAcc	All	FLOAT16 FLOAT32
ElementwiseBaseLayer	Layer to perform Add - Div - Max - Min - Mul operations.	ANEURALNETWORKS_ADD ANEURALNETWORKS_DIV ANEURALNETWORKS_MAXIMUM ANEURALNETWORKS_MINIMUM ANEURALNETWORKS_MUL	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
ElementwiseBinaryLayer	Layer to perform Power and Square Difference operations.	ANEURALNETWORKS_POW	CpuRef	All	FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	FLOAT16 FLOAT32
			GpuAcc	All	FLOAT16 FLOAT32
ElementwiseUnaryLayer	Layer to perform Rsqrt - Exp - Neg - Log - Abs - Sin - Sqrt - Ceil operations.	ANEURALNETWORKS_ABS ANEURALNETWORKS_EXP ANEURALNETWORKS_LOG ANEURALNETWORKS_NEG ANEURALNETWORKS_RSQRT ANEURALNETWORKS_SIN ANEURALNETWORKS_SQRT	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	FLOAT16 FLOAT32 SIGNED32
			GpuAcc	All	FLOAT16 FLOAT32
FakeQuantizationLayer	Layer to quantize float values and dequantize afterwards. The current implementation does not dequantize the values.	N/A	CpuRef	All	FLOAT32
FillLayer	Layer to set the values of a tensor with a given value.	ANEURALNETWORKS_FILL	CpuRef	All	FLOAT16 FLOAT32 SIGNED32
			CpuAcc	All	All
			GpuAcc	All	All
FloorLayer	Layer to round the value to the lowest whole number.	ANEURALNETWORKS_FLOOR	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32
			CpuAcc	All	FLOAT32 FLOAT16
			GpuAcc	All	FLOAT32 FLOAT16
FullyConnectedLayer	Layer to perform a fully connected / dense operation.	ANEURALNETWORKS_FULLY_CONNECTED	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8
			GpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8
GatherLayer	Layer to perform the gather operation along the chosen axis.	ANEURALNETWORKS_GATHER	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	All
			GpuAcc	All	All
GatherNdLayer	Layer to perform the gatherNd operation.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			GpuAcc	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
InputLayer	Special layer used to provide input data to the computational network.	N/A	All	All	All
InstanceNormalizationLayer	Layer to perform an instance normalization on a given axis.	ANEURALNETWORKS_INSTANCE_NORMALIZATION	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32
			CpuAcc	NHWC NCHW	FLOAT16 FLOAT32
			GpuAcc	NHWC NCHW	FLOAT16 FLOAT32
L2NormalizationLayer	Layer to perform an L2 normalization on a given axis.	ANEURALNETWORKS_L2_NORMALIZATION	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	FLOAT16 FLOAT32
			GpuAcc	NHWC NCHW	FLOAT16 FLOAT32
LogSoftmaxLayer	Layer to perform the log softmax activations given logits.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
LogicalBinaryLayer	Layer to perform Logical AND - Logical NOT - Logical OR operations.	ANEURALNETWORKS_LOGICAL_AND ANEURALNETWORKS_LOGICAL_NOT ANEURALNETWORKS_LOGICAL_OR	CpuRef	All	BOOLEAN
			CpuAcc	All	BOOLEAN
			GpuAcc	All	BOOLEAN
LstmLayer	Layer to perform a single time step in a Long Short-Term Memory (LSTM) operation.	ANEURALNETWORKS_LSTM	CpuRef	All	BFLOAT16 FLOAT16 QSYMMS16
			CpuAcc	All	FLOAT16 FLOAT32
			GpuAcc	All	FLOAT16 FLOAT32
MapLayer	Layer to perform map operation on tensor.	N/A	CpuRef	All	All
			CpuAcc	All	All
			GpuAcc	All	All
MaximumLayer	Layer to perform an elementwise maximum of two tensors.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32 SIGNED32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 FLOAT16 FLOAT32 SIGNED32
MeanLayer	Layer to perform reduce mean operation.	ANEURALNETWORKS_MEAN	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
MemCopyLayer	Layer to perform memory copy operation.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 BOOLEAN
			CpuAcc	All	All
			GpuAcc	All	All
MemImportLayer	Layer to perform memory import operation.	N/A	CpuRef	All	All
			CpuAcc	All	All
			GpuAcc	All	All
MergeLayer	Layer to concatenate tensors along a given axis.	ANEURALNETWORKS_CONCATENATION	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
MinimumLayer	Layer to perform an elementwise minimum of two tensors.	ANEURALNETWORKS_MINIMUM	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 FLOAT16 FLOAT32 SIGNED32
MultiplicationLayer	Layer to perform an elementwise multiplication of two tensors.	ANEURALNETWORKS_MUL	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32 SIGNED32
NormalizationLayer	Layer to compute normalization operation.	ANEURALNETWORKS_LOCAL_RESPONSE_NORMALIZATION	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	FLOAT32 FLOAT16
			GpuAcc	NHWC NCHW	FLOAT32 FLOAT16
OutputLayer	A special layer providing access to a user supplied buffer into which the output of a network can be written.	N/A	All	All	All
PadLayer	Layer to pad a tensor.	ANEURALNETWORKS_PAD ANEURALNETWORKS_PAD_V2	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
PermuteLayer	Layer to transpose an ND tensor.	ANEURALNETWORKS_TRANSPOSE	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
Pooling2dLayer	Layer to perform 2D pooling with the specified pooling operation.	ANEURALNETWORKS_AVERAGE_POOL_2D ANEURALNETWORKS_L2_POOL_2D ANEURALNETWORKS_MAX_POOL_2D	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	NHWC NCHW	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
Pooling3dLayer	Layer to perform 3D pooling with the specified pooling operation.	ANEURALNETWORKS_AVERAGE_POOL_3D ANEURALNETWORKS_L2_POOL_3D ANEURALNETWORKS_MAX_POOL_3D	CpuRef	NDHWC	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NA
			GpuAcc	NDHWC
PreCompiledLayer	Opaque layer provided by a backend which provides an executable representation of a subgraph from the original network.	N/A	N/A	N/A	N/A
PreluLayer	Layer to compute the activation layer with the PRELU activation function.	ANEURALNETWORKS_PRELU	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
QLstmLayer	Layer to perform quantized LSTM (Long Short-Term Memory) operation.	ANEURALNETWORKS_QUANTIZED_LSTM ANEURALNETWORKS_QUANTIZED_16BIT_LSTM	CpuRef	All	All
			CpuAcc	All	QASYMMS8 QASYMMU8 SIGNED32 QSYMMS16
			GpuAcc	All	QASYMMS8 QASYMMU8 SIGNED32 QSYMMS16
QuantizeLayer	Layer to perform quantization operation.	ANEURALNETWORKS_QUANTIZE	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 QASYMM16 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QASYMM16 FLOAT16 FLOAT32
QuantizedLstmLayer	Layer to perform quantized LSTM (Long Short-Term Memory) operation.	ANEURALNETWORKS_QUANTIZED_LSTM ANEURALNETWORKS_QUANTIZED_16BIT_LSTM	CpuRef	All	All
			CpuAcc	All	SIGNED32 QASYMMU8 QSYMMS16
			GpuAcc	All	SIGNED32 QASYMMU8 QSYMMS16
RankLayer	Layer to perform a rank operation.	ANEURALNETWORKS_RANK	CpuRef	All	All
			CpuAcc	All	All
			GpuAcc	All	All
ReduceLayer	Layer to perform reduce with the following operations - ARG_IDX_MAX: Index of the max value - ARG_IDX_MIN: Index of the min value - MEAN_SUM: Mean of sum - PROD: Product - SUM_SQUARE: Sum of squares - SUM: Sum - MIN: Min - MAX: Max	ANEURALNETWORKS_REDUCE_MAX ANEURALNETWORKS_REDUCE_MIN ANEURALNETWORKS_REDUCE_SUM ANEURALNETWORKS_REDUCE_PROD	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32 SIGNED32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32 SIGNED32
ReshapeLayer	Layer to reshape a tensor.	ANEURALNETWORKS_RESHAPE ANEURALNETWORKS_SQUEEZE ANEURALNETWORKS_EXPAND_DIMS	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32 BOOLEAN
			CpuAcc	All	All
			GpuAcc	All	All
ResizeLayer	Layer to perform resize of a tensor using one of the interpolation methods: - Bilinear - Nearest Neighbor.	ANEURALNETWORKS_RESIZE_BILINEAR ANEURALNETWORKS_RESIZE_NEAREST_NEIGHBOR	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	NHWC NCHW	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
ReverseV2Layer	Layer to perform reverse of a tensor.	NA	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
RsqrtLayer	Layer to perform Rsqrt operation.	ANEURALNETWORKS_RSQRT	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	FLOAT16 FLOAT32 SIGNED32
			GpuAcc	All	FLOAT16 FLOAT32
ScatterLayer	Layer to scatter updates according to individual values at the specified indices.	N/A	CpuRef	All	FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16 SIGNED32
			CpuAcc	None	N/A
			GpuAcc	All	FLOAT16 FLOAT32 SIGNED32
ShapeLayer	Layer to return the shape of the input tensor.	N/A	CpuRef	All	All
			CpuAcc	All	All
			GpuAcc	All	All
SliceLayer	Layer to perform tensor slicing.	ANEURALNETWORKS_SLICE	CpuRef	All	BFLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
SoftmaxLayer	Layer to perform softmax, log-softmax operation over the specified axis.	ANEURALNETWORKS_LOG_SOFTMAX ANEURALNETWORKS_SOFTMAX	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 FLOAT16 FLOAT32
SpaceToBatchNdLayer	Layer to divide spatial dimensions of the tensor into a grid of blocks and interleaves these blocks with the batch dimension.	ANEURALNETWORKS_SPACE_TO_BATCH_ND	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
SpaceToDepthLayer	Layer to rearrange blocks of spatial data into depth.	ANEURALNETWORKS_SPACE_TO_DEPTH	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All
SplitterLayer	Layer to split a tensor along a given axis.	ANEURALNETWORKS_SPLIT	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
StackLayer	Layer to stack tensors along an axis.	N/A	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
StandInLayer	A layer to represent "unknown" or "unsupported" operations in the input graph. It has a configurable number of input and output slots and an optional name.	N/A	N/A	N/A	N/A
StridedSliceLayer	Layer to extract a strided slice of a tensor.	ANEURALNETWORKS_STRIDED_SLICE	CpuRef	All	BFLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
SubtractionLayer	Layer to perform an elementwise subtract of 2 tensors.	ANEURALNETWORKS_SUB	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16 SIGNED32
			CpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
			GpuAcc	All	QASYMMU8 QASYMMS8 QSYMMS16 SIGNED32 FLOAT16 FLOAT32
TileLayer	Layer to construct a tensor by repeating in tiles a given tensor.	ANEURALNETWORKS_TILE	CpuRef	All	FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16 SIGNED32
			CpuAcc	All	FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QASYMM8 QSYMMS16 SIGNED32
			GpuAcc	All	FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16 SIGNED32
TransposeConvolution2dLayer	Layer to perform 2D transpose convolution (deconvolution) operation.	ANEURALNETWORKS_TRANSPOSE_CONV_2D	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS8 QSYMMS16
			CpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
			GpuAcc	NHWC NCHW	SIGNED32 FLOAT16 FLOAT32 QASYMMU8 QASYMMS8 QUANTIZEDSYMM8PERAXIS
TransposeLayer	Layer to transpose a tensor.	ANEURALNETWORKS_TRANSPOSE	CpuRef	All	BFLOAT16 FLOAT16 FLOAT32 QASYMMS8 QASYMMU8 QSYMMS16
			CpuAcc	All	All
			GpuAcc	All	All
UnidirectionalSquenceLstmLayer	Layer to perform unidirectional sequence LSTM operation.	ANEURALNETWORKS_UNIDIRECTIONAL_SEQUENCE_LSTM	CpuRef	All	Input Types FLOAT32 Weight Types FLOAT32 QASYMMS8
			CpuAcc	All	Input Types FLOAT32 Weight Types FLOAT32
			GpuAcc	All	Input Types FLOAT32 Weight Types FLOAT32
UnmapLayer	Layer to perform unmap operation on tensor.	N/A	CpuRef	All	All
			CpuAcc	NHWC NCHW	All
			GpuAcc	NHWC NCHW	All